Digital Signal Processing
DESD June 2015
M.Kumar
C-DAC, Hyderabad
Agenda
Signal basics
Types of signals, signal classification
Discrete signals, systems, spectrum
LTI system, convolution
ADC(sampling, aliasing, quantization)
Signal processing operations
Basic filters
Introduction
And many more!
GPP Versus DSP
General Purpose Processors(GPP) are primarily designed for data
processing
Primary burden is data read/write
Digital Signal Processors are specifically designed for signal
processing
Handling many tasks at once, controlling huge amounts of memory and
data, controlling a wide variety of computer peripherals (disk drive, modem
etc..)
DSPs are more dedicated, work on smaller range of functions at lightning
speed, yet less costly and requiring much less space (size) and power
Primary burden is mathematical operations
So, DSP architectures incorporate certain features not found in
GPP
Overview of DSP
Signal Processing
Signal processing refers to the science of
analyzing time varying physical processes
Signal Processing
Analog
Before 1970s
Digital
Mixed
1970 to 2000
Current trend
Analog Signal Processing
All real world signals are analog
Uses R,L,C & transistors for signal processing
Analog
Input
Analog
Output
Limitations of ASP
More sensitive to componenet tolerances &
environmental changes
Changes in L, C or power changes system
characteristics
Analog systems take more space
Advantages of DSP over ASP
Ageing has no effect
Flexibility in developing new digital techniques
with minimal hardware changes
Programmable
Uses standard circuits for implementation
Multiplexing is possible
FDM
TDM
Why DSP?
Stability
Programmability and hence flexibility
One H/W can perform several tasks
Upgradeability
More reliable as components are less sensitive
to variations in ambient variables(temp..etc)
Easy to debug, hence short development time
Low cost as independent of H/W complexity
DSP Applications
DSP systems must work in real time, capturing
and processing information as it happens.
Otherwise, information is lost and the signal
gets distorted
Calculations should be done as fast as the
signal data is received from the ADC
DSP in Real World
DSP-based VoIP gateway makes a shift to data
networks possible, serving as the bridge between
PSTN and the packet network
Speech compression & transmission in mobile phones
Image, audio, video processing
Medical Imaging
Weather forcasting
Seismic data processing
Control of industrial processes
DSP Processors
TMS 320 from texas Instruments
Motorola DSP 56000
SHARC processors from Analog Devices
Recent Advancements in DSPs
Run at horse speeds
Smaller footprint
Reduction in power dissipation
DSP Application
Source:
1. http://news.bbc.co.uk/2/hi/health/8521067.stm
2. http://www.matchboxmobile.com/
What's Inside a DSP?
Signal Basics
What is a signal?
A signal is a function of one or more independent
variables such as time, distance, temparature etc
signal = function(x1,x2,x3...)
Why signals?
Signals carry useful information
Why signal processing?
To extract useful information carried by the signal
Speech Signal
Biological Signals
Biological signals are time series signals
generated by some biological mechanism
ECG(Electrocardiography) signals are electrical
potentials recorded from the chest(mainly),
arms, legs
EEG(Electroencephalography) signals are
oscillating electrical potentials recorded from
the scalf surface
They are generated by neuronal activities in the
brain
Example
Research
'Mind-reading machine' can convert thoughts
into speech
Source: http://mashable.com/2011/01/02/real-sci-fi-tech/#6643MindReading-Machine
http://www.eurekalert.org/pub_releases/2010-09/uou-tbs090110.php
Mind Reading Machine
This photo shows two kinds of electrodes
sitting atop a severely epileptic patient's
brain after part of his skull was removed
temporarily
An array of 16 microelectrodes known as a
microECoG grid is arranged in a four-by-four
array and shown next to a US quarter-dollar coin
with a Utah state
This magnetic resonance image (MRI) of an epileptic patient's
brain is superimposed with the locations of two kinds of
electrodes: conventional ECoG electrodes (yellow) to help locate
the source of...
i-Brain
Image Signal
Light intensity as a function of 2D coordinates
B&W(0 or 1) or grey scale images(I=0-255)
Color images(I=R(0-255),G(0-255),B(0-255))
Video Signal
Sequence of images called frames
Video is a function to three variables(2 spatial
coordinates and time)
Classification of Signals
Signals can be classified into various types by
Nature of independent variable
Value of the function defining the signal
Discrete/Analog(Continous)
Deterministic/Random
Periodic/Aperiodic
Depending on the dimensions
Even/Odd etc...
Classification based on Amplitude and Time
1.Analog
2.Discrete
3.Digital
4.Quantised boxcar
Amplitude- continuous
Time-continuous
Amplitude- continuous
Time-discrete
Amplitude- discrete
Amplitude- discrete
Time-discrete
Time-continuous
Discrete Vs Continous
Normally, independent variable is time
Continous time signal
Time is continous
Defined at every instant of time
Discrete time signal
Time is discrete
Defined at discrete instants of time- it is a sequence
of numbers
Random Vs Deterministic Signal
Deterministic
A signal can be predicted using some mathematical
expression, rule or look up table
Behavior is fixed and easy to analyse
Future values of the signal can be calculated from
past values
An example of a deterministic signal, the sine wave
Random (or stochastic)
A signal is generated randomly and can not be
predicted ahead of time
Future values can not be predicted and it is difficult
to analyse
Probability and statistics are used to analyse their
behavior
Because of their randomness, average values from
a collection of signals are usually studied rather
than analyzing one individual signal
An example of a Random signal, the sine wave with noise added
Periodic Vs Aperiodic
A continous time signal is periodic if it exhibits
periodicity, x(t+T)=x(t), -<t<, where T=period
A periodic signal has a definite pattern that
repeats over and over with a repetition period of
T0
Periodic signal (continuous-time)
For discrete time signal, x(n+N0) = x(n), -<n<
A signal, which does not have a repetitive pattern is aperiodic
Even/Odd Signal
Even
Signal exhibits symmetry in time domain
x(t) = x(-t) or x(n) = x(-n)
Odd
Signal exhibits anti-symmetry in the time domain
x(t) = -x(-t) or x(n)=-x(-n)
A signal can be expressed as sum of even and odd
components x(t) = xeven(t) + xodd(t)
where xeven(t) = 1/2[x(t)+x(-t)], xodd(t) = 1/2[x(t)-x(-t)]
Dimensions of a signal
A signal can have any number of dimensions
1D signal = function(x); where x = time, distance etc
2D signal = function(x,y); where x,y = spatial
positions
3D signal = function(x,y,z); where x,y,z = spatial
positions
Most signals are generated naturally but they
can also be generated artificially using a
computer simulation (Matlab/Octave)
Singular Functions
Important non-periodic functions
Delta or unit impluse function is the most basic
function
All other singular functions can be derived out
of it
t =0, t0 ; t dt=1
(n)=1, n=0
=0, n0
Unit step function
u(t)=0,t<0
=1,t>0
u(n)=0, n<0
=1,n>=0
Unit ramp function
r(t) = 0, t<0
= t, t>0
r(n) = 0,n<0
= n, n>=0
Unit pulse function
t =u t+
1
1
u t
2
2
Why Discrete Time Signals?
Most of the signals found in real world are
continous in nature
In order to process these signals, we need to
convert them to a convenient form so that we
can apply digital techniques
Analog is used to describe a waveform that's
continous in time and can take on a continous
range of amplitude values
Example
Voltage applied to oscilloscope
Discrete time signal is used to describe a signal
whose independent time variable is quantised
so that we know the value of the signal at
discrete instants of time
Discrete Time Signal
In addition to quantizing time, a discrete time
signal quantizes the signal amplitude
Example
Time Domain Signals
Discrete System
A discrete system is a collection of h/w or s/w
routines, that operate on discrete time signals
Example
If x[n] = {1,3,5,7,9} then y[n] = ?
How to measure frequency of continous
sinewave?
Ans: Apply it to an oscilloscope, spectrum
analyser or frequency counter
Given 20 signal samples, can we measure
frequency of discrete sinewave?
Answer
No, we can just plot the samples
However, given sampling period ts, we can know
the frequency of discrete sinewave
Example: If ts = 0.05 ms/sample,
Period of sinewave = 20 samples/period x 0.05
ms/sample = 1 ms
Frequency = 1/period = 1/1ms = 1kHz
So, in discrete systems, absolute frequency
determination is dependent on sampling
frequency fs =1/ts
Frequency Domain Representation
Discrete sinewave x1(n) with an arbitrary
frequency f0 Hz can be represented as
The same x1(n) can be described by indicating
that it has a frequency of 1, measured in units
of f0 , and no other requency content
Example
Summary
A discrete time domain can be represented
Mathematically as x1(n)
Graphically
Discrete frequency domain X1(m)
Signal Amplitude
The amplitude of a variable is the measure of
how far, and in what direction, that variable
differs from zero
So, signal amplitude can be +ve or -ve
Signal Magnitude
The magnitude(absolute value) of a variable is
the measure of how far , regardless of
direction, its quantity differs from zero
So, magnitudes are always +ve
Signal Power
The power of a signal is proportional to its
amplitude or magnitude squared
Signal Processing Operations
Adder: element for element addition of two
different sequences, to produce a new
sequence
Subtractor
Summation
Multiplication
Unit Delay
Example: a(5) = b(4) etc...
Summary
Signal processing operations are always
performed on sequences of individual discrete
values
The elementary operations are very simple
Regardless of how complicated they appear to
be, most of the DSP algorithms can be
performed using the combinations of these
simple operations
LTI Systems
Why LTI Systems?
Answer
Linearity and time invariance are two important
system characteristics
Most of the real world systems are LTI
It is easy to predict performance of any DSP
scheme, as long as it is LTI
Linear
Linear defines a special class of systems where
the output is the superposition or sum of the
individual outputs had the individual inputs
been appplied separately to the system
If x1(n) -----------> y1(n),
x2(n) -------------> y2(n) then
x1(n) + x2(n) -------------> y1(n) + y2(n)
Proportionality Characteristic
If the inputs are scaled by constant factors c1
and c2 then the output sequence parts are also
scaled by those factors
c1x1(n) + c2x2(n) -------------> c1y1(n) + c2y2(n)
This is also known as homogeneity property
Input x(n)
Linear Discrete
System
Output y(n)
Example
Lets take a discrete system whose
y(n) = -x(n)/2
Input x(n)
Linear Discrete
System
Output y(n) = -x(n)/2
Let x1(n) represent 1Hz sinewave sampled at a
rate of 32 samples per cycle and x 2(n)
represent 3Hz sinewave
Non-Linear System
y(n) = [x(n)]2
Input x(n)
Linear Discrete
System
Output y(n) = [x(n)]2
Let x1(n) = 1 Hz sinewave and x2(n) = 3 Hz sinewave
x1(n) = sin(2f0nts) = sin(2.1.nts)
y1(n) = [x(n)]2 = sin(2.1.nts).sin(2.1.nts)
using
Therefore, y1(n) has 0Hz and 2Hz frequency components
Similarly, when x2(n) is applied, y2(n) will contain
0Hz and 6Hz components
When x3(n) = x1(n) + x2(n) is applied, we can
predict the freq content of y3(n) output sequence
by using (a+b)2 = a2 + 2ab +b2, where a and b
represent 1Hz and 3Hz sinewaves respectively
a2 generates 0 and 2 Hz sinusoids, b 2 generates
0 nd 6Hz sinusoids, and 2ab yields another 2Hz
and 4Hz sinusoids in y3(n)
These extra sinusoids were generated by an interaction of
the two input sinusoids due to the squaring operation. That's
Non-Linearity
Summary
The linear systems output is the sum of the
individual outputs
Also, the output y1(n) depends only on x1(n), not
on other input x2(n)
There is no interaction between inputs x 1(n) and
x2(n) at the output of linear systems
Time-Invariant Systems
A time-invariant system is one where a time
delay(or shift) in the input sequence causes a
equivalent time delay in the system's output
sequence
If x(n)---------> y(n) then
x'(n) = x(n+k) -------------> y(n+k)
where k is some integer representing k sample
period time delays
Example
Let x(n) is unit amplitude, 1Hz sinewave
sequence with y(n) as output
Input x(n)
Linear Discrete
System
Output y(n) = -x(n)/2
Let x'(n) = x(n+4), then x'(0) = x(4), x'(1)=x(5)...
then y'(n) = y(n+4)
i.e..y'(0) = y(4), y'(1) = y(5).....
Commutative Property of LTI system
Analyzing LTI Systems
LTI systems can be analysed to predict their
performance
If we know the unit impulse response of an LTI
system, we can calculate everything there is to
know about the system
The LTI system completely characterizes the
system
Input x(n)
LTI
Discrete System
Output y(n)
Impulse Response
Impulse response means system's time domain
output sequence when the input is a single
unity-valued sample(unit impulse) preceeded
and followed by zero valued samples
Adantages of LTI System
We can determine the output sequence for any
input sequence
output seq = convolution of input sequence and
system's impulse sequence
Also, we can find the system's frequency
response by taking the discrete fourier
transform of that impulse renponse
Analog to Digital Conversion
Analog to Digital Conversion
Analog
Analog
Digital to Analog Conversion
Periodic Sampling
Periodic sampling is the process of representing a
continous signal with a sequence of discrete data
values
Continous Signal
Analog to Digital
Converter
Digital Values
Question:How fast a given continous signal must be
sampled in order to preserve its information content?
Frequency Ambiquity
There is a frequency ambiquity associated with
dicrete time signal samples
Example
If we were given the following sequence of
values x(0) =0, x(1)=0.866, x(2) = 0.866,
x(3)=0, x(4)=-0.866, x(5)=-0.866, x(6)=0
Next, two people were asked to draw that
sinewave
Samples
Two sinewaves passing through the same
points
The key issue is, if data sequence represented
periodic samples of a sinewave, we can not
unambiguously determine the frequency of the
sinewave from those sample values alone
Let x(t) = sin(2f0t) and
x(n) = sin(2f0nts) where ts = 1/fs
Since
x(n) = sin(2f0nts) = sin(2f0nts + 2m)
= sin(2(f0 + m/nts))nts
If we let m be an integer multiple of n, m=kn, we
can replace m/n with k, so that
x(n) = sin(2f0nts) = sin(2(f0 + kfs)nts)
Here, the f0 and (f0 + kfs) are therefore equal
Important Points
x(n) sequence of digital sample values,
representing a sinewave of f0 Hz, also exactly
represent sinewaves at other frequencies
namely f0 + kfs
When sampling at a rate of fs samples, if k is
any positive or negative integer, we cannot
distinguish between the sample values of a
sinewave of f0 Hz and a sinewave of (f0 + kfs )
Hz
The sample nature of any sequence of discrete
values makes that sequence also represent an
infinite number of different sinusiods
Also, from equation-1, the spectrum of any
discrete series of sampled values contains
periodic replicas of the original continous
spectrum
The period between these replicated spectra in
the frequency domain will always be fs
These spectral replications repeat in both
directions of frequency spectrum. That's
because k can be any +ve or -ve integer
Also, equation-1 is the reason that all digital
filter frequency responses are periodic in the
frequency domain
Example
Consider the sampling of a 7kHz sinewave at a
sampling rate of 6kHz
Equation-1 is causing a sinusoid whose name
is 7kHz, to go by the alias of 1kHz
Notice that the sample values would not
change at all if, instead, we were sampling a 1
kHz sinewave
Here f0=7kHz, fs=6kHz and k=-1 such that
f0 + kfs = [7 + (-1).(6)]=1kHz
Our problem is that no processing scheme can
determine if the sequence of samped values,
whose amplitudes are represented by the dots
came from a 7kHz or a 1kHz sinusoid
Another Example
Here f0=4kHz, fs=6kHz and k=-1 such that
f0 + kfs = [4 + (-1).(6)]=-2kHz
Now, we restrict our spectral band of interest to
the frequency range of fs/2 Hz
In sampling theory, fs/2 is called by different
names as critical Nyquist, Half Nyquist and
folding frequency
Within the spectral band of interest(3Hz
because fs=6Hz), there is energy at -2 kHz and
+1 kHz, aliased from 4kHz and 7kHz
respectively
Note, here that the vertical positions of the dots
in the figure have no amplitude significance
Only, horizontal positions indicate which
frequencies are related through aliasing
Sampling Low Pass Signals
A band limited signal implies that any signal
energy outside the range of BHz is below the
sensitivity of our system
This continous spectrum can't be represented
in a digital machine in its current band limited
form
If we change fs through decimation or
interpolation, the spectrum replicas will shift up
or down right in the middle of the frequency
range of interest fs/2, which causes problems
How can we control the locations of those
spectral replications?
Answer: Using Nyquist Criteria
Nyquist Criteria
In practical ADC, the fs 2B is used to separate
spectral replications at the folding frequencies
of fs/2
Lets take fs = 1.5B, which results in
undersampling
Here the spectral replications are now
overlapping the original baseband spectrum
centered about zero Hz
Observation-1
Limiting our attention to the band of fs/2 Hz,
we see two important points
The lower edge and the upper edge of the
spectral replicas centered at +f s and -fs now lie
in our band of interest
Portions of spectral replications now combine
with the original spectrum which results in
aliasing errors
Aliasing
Now, the discrete sampled values associated
with the spectrum of fig-3 no longer truly
represent the original input signal
The spectral information between -B to -B/2
and B/2 to B Hz has been corrupted
We can not know the exact amplitude in the
aliased region
Observation-2
The entire spectral content of the original
continous signal is now residing in the band of
interest between -fs/2 and +fs/2
Also, any signal energy located above +B Hz
and -B Hz in the original continous spectrum
will always end up in the band of interest after
sampling, regardless of the sampling rate
This property is true regardless of the original
signal or the sample rate
Anti Aliasing Filter
That's the reason why all the continous signals
are band limited by passing through analog low
pass filters before applying to ADC
Example
Consider a continous signal of bandwidth B
accompanied by noise energy
Sampling the continous signal with fs>2B
prevents replications of the signal of interest
from overlapping each other, but all of the noise
energy still ends up in the range of -f s/2 and
+fs/2 of our discrete spectrum
So, use anti-aliasing filter prior to ADC
conversion to attenuate any unwanted signal
energy above +B and below -B Hz
Original
Continous
Signal
Analog Anti
Aliasing Filter
(cutoff freq=BHz)
Filtered
Continous
Signal
Discrete
Samples
ADC
Binary Quantization
To process a sampled signal in a computer,
samples must be converted to binary numbers
using a finite number of bits per sample
N bits per sample gives 2N quantization levels
More bits per sample(larger N) means better
resolution, but the digital data becomes larger
and results in complex DSP architecture
Quantization involves rounding operations.
Therefore the accuracy of the quantized signal is
limited
The round off error induced by quantization is
called Quantization Noise
The maximun quantization error per sample is
equal to the half the binary Step Size
Difference between quantized and original
value is called quantization error or noise
Spacing between two quantization levels is
called quantization width or resolution
For quantization to N bits, the quantization error
Qe of each digital sample can be viewed as the
uniformly distributed random variable in the
range of -2-N < Qe < 2N
For a B bit quantizer, SQNR = 6B dB
For every additional bit, SQNR has a gain of
6dB
More bits per sample(larger N), gives smaller
quantization error
However, a DSP system with more bits per
sample becomes expensive
Example
Resolution = Range/Number of Levels(2 N)
For a signal varying between 0 to 5V with a 8
bit ADC has a resolution of ____and SQNR___
Summary
Basic Concepts
Types of Signals
Classification of Signals
Discrete Signals,
Systems, Spectrum
LTI Systems
ADC, Sampling, Aliasing
Signal
Analog
Continuous
Time
Digital
Discrete
Time
ADC
How to process a signal?
Depends on the type of the signal and the
information being carried by the signal
Type of the signal can be in terms of basis
functions in the domain of original independent
variable(s) or the transformed domain
Also, Information extraction process may be
carried out in the original domain of the signal
or the transformed domain
Signal processing Operations
Analog processing
Discrete processing
Mostly time domain operations
Both time & frequency domain operations
Filtering is a widely used signal processing
operation
Simple Time Domain Operations
Scaling: Multiplication of signal with +ve or -ve
constant, y(t) = kx(t)
Where amplification if k>1, attenuation if k<1
Delay: Delayed replica of original signal
y(t) = x(t-t0)
Addition: Two or more signals combine to
generate a new signal, y(t) = x 1(t) + x2(t)
Multiplication, Integration, Differenciation etc...
Signal Processing Examples
Filtering
Removing unwanted signal from the desired signal
Multiplexing
More signals are combined and transmitted over
the same channel
Filtering
An important frequency domain operation
A filter passes certain frequency components
with minimal distortion and blocks all other
frequency components
Basic filters are
LPF
HPF
BPF
BSF
Thank you
[email protected]
FDM Vs FMDA
FDM is a technique that is used in many other
technologies. A multiplexer modulates all the
signals that is going to use the channel into a
single signal
FDMA foregoes the use of a multiplexer, since it
acts on the data link layer. All the information is
combined prior to generating a signal, making
the use of a physical layer multiplexer
unnecessary
Network Layers
Case Study
Low Pass Filter
1.5
1.5
1
LPF
3KHz
0.5
2KHz
0
-0.5
0
-0.5
-1
-1
-1.5
-1.5
200
400
600
800
1000
200
400
600
800
1000
800
1000
0.5
0
0
-1
-0.5
-2
-1
-3
-1.5
Only 2 Hz signal
remains
1.5
5KHz
0.5
0
0
200
400
600
800
200
400
600
800
1000
1000
3
2
1
1.5
LPF
8KHz
1
0.5
11KHz
-2
-3
0
-0.5
-1
-1.5
0
-1
200
400
600
800
1000
200
400
600
Only 2 Hz and 5 Hz
signals remain
High Pass Filter
1.5
0.5
2KHz
HPF
3KHz
0
-0.5
-1
-1.5
0
200
400
600
800
-2
1000
200
400
600
800
1000
Only 5 Hz and 11 Hz
signals remain
1.5
0.5
-1
-0.5
-2
-1
-3
-1.5
0
0
-1
-3
5KHz
0
200
400
600
800
1000
200
400
600
800
1000
Combined signal
1.5
1
1.5
HPF
8KHz
1
0.5
11KHz
0.5
0
-0.5
-1
-0.5
-1.5
-1
-1.5
0
200
400
600
800
1000
200
400
600
800
1000
Only 11 Hz signal remains
Band Pass Filter
1.5
1
0.5
2KHz
0
-0.5
-1
-1.5
0
200
400
600
800
1000
1.5
+
1.5
1
0.5
BPF
-1
-2
-1
-3
200
400
200
400
600
800
1000
600
800
1000
1
0.5
0
-0.5
-1
-0.5
-1
-1.5
1.5
-1.5
0
-0.5
-1.5
0
11KHz
0.5
5KHz
200
400
600
800
Combined signal
1000
fp1=4 Hz,
fp2=6 Hz,
fs1=3 Hz,
fs2=7 Hz
200
400
600
800
Only 5 Hz signal
remains
1000
Band Stop Filter
1.5
1
0.5
2KHz
0
-0.5
-1
-1.5
200
400
600
800
1000
1.5
+
1.5
1
0.5
BSF
-1
-0.5
-2
-1
-3
0
-0.5
-1
-1.5
200
400
600
800
1000
1
0.5
0
-0.5
-1
-1.5
-1.5
0
0
1.5
11KHz
0.5
5KHz
200
400
600
800
1000
200
400
600
800
Combined signal
1000
fp1=4 Hz,
fp2=6 Hz,
fs1=3 Hz,
fs2=7 Hz
200
400
600
800
1000
5 Hz signal is filtered
out, only 2 Hz and 11
Hz signals remain
Case Study
Shoot
Capture
Play Back
Edit
Serve
Encode
Integrate
